There are two ways of going broke with GPU computing. One is by buying a card that you can't afford to pay, the other is by buying a card which power needs are such that you can't afford to pay the power-bill.
Just as with people that write about cars, who have the tendency to describe the latest Ferrari's, Lamborghini's and other unobtainable sports cars, I will now describe which nVidia-cards to avoid like the plague, however good performers they are. They will let you go broke both ways!
*=prosumers only
Model/Type GPU Fab in nm Shaders TMU ROP GFLOP SP GFLOP DP TDP Gflop(SP)
/WattGflop(DP)
/WattTesla 1/8 of SP GTX 295 2x GT200-400-B3 55 2x 240 2x 80 2x 28 1192 149 289 4.12 0.52 Quadro Plex 2200 S4 4× GT200GL 55 4x 240 4x 80 4x 28 2488 311 1200! 2.07 0.26Fermi 1/8 of SP* GTX 590 2× GF110-351-A1 40 2x 512 2x 64 2x 48 2488 311 365 6.82 0.85 S2070 GPU Computing Server 4× GF100 40 4x 448 4x 56 4x 40 4120 2060 900 4.56 2.28Kepler 1/24 of SP* GeForce GTX Titan Z 2× GK110 28 2x 2880 2x 240 2x 48 8122 2707 375 21.66 7.22 K80 GPU Accelerator 2× GK210 28 2x 2496 2x ? 2x ? 8736 2912 300 29.12 9.71Maxwell 1/24 of SP* GeForce GTX Titan X GM200-400 28 3072 192 96 6144 192 250 24.58 0.77 M60 GPU Accelerator 2× GM204 28 2x 2048 2x ? 2x ? 9650 302 300 32.17 1.01Pascal 1/32 of SP* TITAN XP GP102 16 3840 240 96 12150 380 250 48.60 1.52 P100 GPU Accelerator GP100 16 3584 ? ? 10609 5304 300 35.36 17.68